Comment by davebren

It has. LLMs being able to one-shot any higher-order complexity is entirely dependent on it already being in the training data.